feathub
feathub copied to clipboard
alibaba
FeatHub - A stream-batch unified feature store for real-time machine learning
local_ast_evaluator.py ```python def eval_is_op(self, ast: IsOp, variables: Optional[Dict]) -> Any: left_value = self.eval(ast.left_child, variables) is_none = left_value is None # Pandas converts None value of numeric type to NAN. Therefore,...
```python # PyFlink Table#to_pandas currently doesn't support Map type. We have to collect the # result and construct the pandas DataFrame. # TODO: Use PyFlink Table#to_pandas after # https://issues.apache.org/jira/projects/FLINK/issues/FLINK-30607 is...
`flink/table_builder/source_sink_utils_common.define_watermark` should properly handle timestamp with or without timezone when `timestamp_format == 'epoch'`
figure out why Flink SQL requires a processing time attribute for lookup join and see if there is space for improvement on Flink API.
In `flink_table_builder.py`, the method `_apply_filter_if_any` is applied after transformations. This filtering should be applied before transformations and other logics if possible to reduce computations. Same for `spark_dataframe_builder.py`.
PyFlink 1.16 and 1.17 have problem working with multiple StreamTableEnvironment. Creating a StreamTableEnvironment will overwrite the context class loader that belongs to the previous StreamTableEnvironment, which can cause class loading...
Change the default value of collection-typed agg functions (COLLECT_LIST, VALUE_COUNTS, etc) to empty collection after FLINK-32494 is resolved
remove this method as it's function can be supported by Java's AggFunc.getResult(AggFunc.createAccumulator()).
Blocked by flink ticket: FLINK-27297